How We Turned Generic AI Into a Specialist — And What That Means for Your Business
Most businesses get mediocre AI output because they ask AI to think and create in a single step. Building a production AI pipeline with over 1,000 lines of carefully chosen prompting revealed a better approach — and the principles apply to any business using AI.
Most people experience AI like this: you type in a prompt, it gives you something back, and it is… fine. Competent, generic, and just good enough that you spend twenty minutes editing it into something your team would actually use.
I had the same experience. So I spent months fixing it.
The result was DreamCopy — a production system I built that takes up to 30 property photos, layers in location and market intelligence, and produces four distinct, publication-grade real estate listings that agents use with minimal or no editing. It runs end-to-end without human intervention. Over 1,000 lines of prompting, every word carefully chosen.
But this is not a product story. It is an architecture story — about what I learned building it, and why those lessons apply to any business that wants AI output to be specific, reliable, and on-brand, not “generic but fine.”
The core insight: the system you build around AI matters more than the model itself.
From “Write Me a Listing” to “Understand This Asset First”
If you ask an AI model to “write a listing for a 4-bedroom coastal home,” you will get exactly what you would expect. Something about “stunning ocean views” and “entertainer’s kitchen” and “sought-after location.” It sounds like every other listing because, from the AI’s perspective, this property looks like every other property.
The instinctive fix is to stuff more detail into the prompt. Describe every feature yourself and ask the model to rewrite it. At that point, the AI is not really thinking. It is a clumsy word-processor.
I took a different approach.
Instead of asking AI to “write a listing,” DreamCopy first asks it to understand the asset through a structured, multi-step process — then write. The pipeline has twelve stages. The first seven are pure analysis. The next four generate content. The last one edits. Each step builds on the previous, compounding understanding in a way a single prompt never can.
In the property context, that means the AI does not just label “kitchen” and “bedroom.” It looks at how spaces connect, how light falls, what daily life would feel like in that home, who the likely buyer is, and how the property sits in its local market. Only then does it start writing.
For a business owner, the analogy is simple: do not ask AI to “write the proposal” or “draft the product page” from a thin prompt. Make it understand the asset, the client, and the context first — then create.
What I Actually Built (In Plain English)
Under the hood, DreamCopy does three things that generalise to almost any AI use case.
1. Separate analysis from creation
The work is split into distinct thinking stages:
- The system first extracts and structures facts from inputs (photos, address, agent notes, specifications).
- It then analyses and interprets those facts: what matters, to whom, and why.
- It synthesises everything into a single, coherent understanding of the property.
- Only after that does it generate four different listing styles, each for a specific channel or purpose.
In practice, that means seven analytical steps before a single word of copy is written. The writing is better because the understanding is better.
For your business: if AI output feels generic, you are probably asking it to think and create in one shot. Break the job into stages — extract, analyse, synthesise, then generate.
2. Turn vague context into usable intelligence
An address like “Eagle Bay” or “Cottesloe” is more than a pin on a map. In DreamCopy, the system translates location into lifestyle advantages, neighbourhood character, likely buyer profile, and market context. That turns “nice coastal suburb” into specific positioning: who this is for, what they value, and what they will pay a premium for.
For your business: whatever your “location” is — client industry, market segment, product category — treat it as data to be unpacked, not a label. Build steps that explicitly translate it into advantages, risks, and positioning.
3. Build one shared source of truth for the AI
At the midpoint of the pipeline, the system stops and answers a single question: what is the core story here?
I use a simple hierarchy:
- Hero benefit — the experience or outcome that matters most to the buyer.
- Hero feature — what delivers that outcome.
- Solution — how it is actually executed.
- Proof — concrete, visible evidence.
Each layer must connect to the next with the word “because.” If the chain breaks, the story is not grounded enough to use.
A property example: “Morning light fills the living space all day” — because — “floor-to-ceiling north-facing glazing” — because — “double-height window wall spanning the full width” — because — “visible in photos 3, 7, and 12.”
The same structure works for any product or service. A consulting firm: “Clients act on recommendations immediately” — because — “every finding is prioritised with expected ROI” — because — “the evaluation framework quantifies impact, not just identifies issues” — because — “20 years of writing reports that went to people who were accountable for results.”
This hierarchy becomes a single synthesis object that every later step can see. Whether the AI is writing a long editorial listing, a bullet-point portal description, or a concise social post, it draws from the same underlying understanding rather than reinventing the story each time.
For your business: before you ask AI to create ten versions of something, make it agree with you on the one core narrative first — then have it express that narrative in different formats.
The Design Rules That Emerged
Months of iteration on DreamCopy surfaced a handful of rules I now apply across every AI project.
Rule 1: Tell AI what to do, not what to avoid
Early prompts were full of negatives: “Do not invent features. Do not be generic. Avoid hype.” The result was cautious, hedged, and bland.
I changed every “do not” into a positive requirement:
- “Do not invent features” became “Use only features found in the data provided.”
- “Do not be generic” became “Use specific, concrete terminology.”
- “Avoid passive voice” became “Use active voice exclusively. Every sentence must have a clear subject.”
- “Do not hype” became “Maintain quiet confidence. No exaggeration.”
The tone shifted immediately. The model stopped writing like a compliance document and started writing like a confident specialist.
For your business: audit your prompts. Wherever you see “do not” or “avoid,” flip it into a clear instruction about what you do want. (I wrote a complete guide to this technique with practical examples.)
Rule 2: Match the creativity dial to the job
DreamCopy uses different temperature settings for different tasks:
- Very low for fact extraction and data formatting — minimise hallucination.
- Moderate for strategic interpretation and positioning — allow creative reading of the landscape.
- Slightly higher for narrative writing — balance consistency with natural language flow.
- Low again for editing — polish without invention.
Most setups use one setting for everything, which either produces hallucinated “facts” or flat, lifeless prose. Different cognitive tasks need different tolerances for creativity.
For your business: if you are running multi-step workflows, deliberately tune how much freedom the AI has at each stage instead of leaving one global default.
Rule 3: Use two editorial roles, not one
The final polish step only started working when I gave the AI two jobs at once:
- A “writer’s advocate” whose job is to protect voice, emotional impact, and buyer desire.
- A hard-nosed editor whose job is to cut anything that does not add evidence, clarity, or differentiation.
The editor’s core instruction: “Strengthen by subtraction. When in doubt, delete rather than replace. Never add new claims or embellishments.”
That structured tension produces copy that is both evocative and tight. One role alone either over-protects or over-cuts.
For your business: when refining AI output, consider designing prompts with two explicit perspectives — for example, “client advocate” and “compliance editor” — and make them work together on the same draft.
Rule 4: Let context compound
Every generative step in DreamCopy receives the full synthesis: rooms, themes, buyer profile, location intelligence, and positioning. Even when writing a simple closing paragraph, the system still “knows” the quality of the kitchen or the character of the neighbourhood. That knowledge shapes the confidence and specificity of the language, even in sections that never mention those details directly.
I tested trimming context to save cost. Output quality dropped in ways humans could feel, even when the model technically “worked.”
For your business: give AI rich, repeated context for important tasks. The extra tokens are usually trivial compared with the value of output your team can use without rewriting.
The Uncomfortable Truth About AI Quality
There is a truth most AI vendors gloss over: the model is rarely the limiting factor. The same underlying model that produces generic text from a single prompt can produce publication-ready copy when it sits inside a well-designed pipeline.
Upgrading to the latest, most expensive model will not fix a weak architecture. It just gives you slightly better-sounding generic output.
If your AI experiments are producing work your team does not trust or does not use, the solution is probably not “more model.” It is a better system around the model.
The Bottom Line
The difference between AI that produces “a draft you fix” and AI that produces “a draft you ship” is not the model. It is the architecture.
Separate analysis from creation. Build a single source of truth before generating anything. Give the AI a clear identity through positive framing, not a cage of prohibitions. Match creativity settings to the cognitive task. Use competing editorial roles for refinement. And let context compound rather than fragmenting it for efficiency.
These are not theoretical principles. They are the product of months of production iteration, thousands of generated outputs, and over 1,000 lines of prompting where every word was chosen for a reason.
They apply to property listings. They apply to proposals, product pages, reports, and marketing copy. They apply anywhere you need AI to produce work that is specific enough, reliable enough, and good enough that your team uses it without reaching for the red pen.
The gap between businesses that prompt AI casually and businesses that build structured AI workflows is widening. It will become competitive advantage. Not because of the model — but because of the architecture around it.
At Perth AI Consulting, this is the work I do: design the architecture around AI, not just the prompt. If you are exploring AI for your business and want output that feels like it came from your best people, start with a conversation.